Fix speech editing boundary artifacts by working in mel domain

[whk-sjtu]

Previously, speech_edit.py worked in wav domain (inserting zeros into the waveform before computing mel spectrogram), which caused boundary artifacts when mel spectrogram windows straddled zeros and real audio.

This commit refactors the approach to work in mel domain:

• Compute mel spectrogram on the clean original audio first
• Insert zero frames in mel domain instead of zero samples in wav domain
• Use frame-level granularity throughout for consistency

Benefits:

• Eliminates boundary artifacts
• More consistent behavior regardless of small float variations in input times
• Cleaner edit boundaries

Changes to speech_edit.py (lines 148-220):

• Convert audio to mel using model.mel_spec() before editing
• Build mel_cond by concatenating original mel frames + zero frames
• Calculate all time-based values at frame level first, then convert to samples
• Pass mel_cond directly to model.sample() instead of raw audio